Qwen3.8 27B
on M5 Pro 64GB · Inco Splash · 88,535 ctx
Use cases
tool-use
Summary
User benchmarks Qwen3.8-27B at 76.2 tok/s decode and 392.6 tok/s prefill on an M5 Pro with 64 GB, using Inco Splash 1.5.0 with a Jundot oQ4e quant and DFlash2 speculative decoding.
The run used an 88,535-token prompt including a system message and three tool schemas, with reasoning_effort=medium, temperature 0 and seed 17; time to first token was 225.51 s and output was capped at 2,048 tokens.
The post is the first row of a 12-engine comparison; the user notes it is not a scientific comparison since models and quantizations may differ between engines.